基于半监督学习的三维Mesh建筑物立面提取与语义分割方法

Semi-supervised Learning Based 3D Mesh Building Facade Extraction and Semantic Segmentation Method

面向三维Mesh数据的建筑物立面自动提取与语义分割在智慧城市建模与分析、数字孪生城市、城市规划建设等领域中应用广泛。拟基于深度学习方法构建面向三维Mesh数据的建筑物立面语义分割模型。当前,大量基于卷积神经网络的深度学习模型均是全监督学习方法,其性能严重依赖于人工标注的训练集质量。但高质量的Mesh三维场景人工标注数据集昂贵且稀缺,鉴于此,在自动提取建筑物立面数据的基础上,提出一种基于mean-teacher半监督学习的三维建筑物立面语义分割方法,并引入特征空间关系正则化,结合空间和特征方面的邻域结构,利用无标签数据来提升模型分类精度。构建了一个全新的基于三维Mesh数据的建筑物立面数据集,并通过实验验证了提出方法的有效性和可用性。

Automatic extraction and semantic segmentation of building facades for 3D Mesh data was widely used in the fields of smart city modeling and analysis,digital twin city applications,and urban planning and construction.A semantic segmentation model of building facades for 3D Mesh data based on deep learning methods was proposed.Currently,a large number of deep learning approaches based on Convolutional Neural Networks(CNN)were fully supervised,and their performance depended heavily on the quality of the manually labeled training date set.However,high-quality manual annotated datasets of mesh 3D scenes were expensive and scarce.In view of this,a Mesh 3D building facades automatic extraction approach was designed,a semantic segmentation method for 3D building facades based on mean-teacher semi-supervised learning was proposed.In order to enhance the performance of our method further,the feature spatial relationship regularization method(FESTA)was introduced,which combined the neighborhood structure in terms of space and features,and used unlabeled data to improve the model classification accuracy.Finally,the effectiveness and usability of the proposed method were experimentally verified.